Set up your AI: accounts, models, privacy, charts

1Pick one main tool

Claude, ChatGPT and Gemini can all read images, follow standing instructions and keep to an output format. Any one of them can run this team. Pick the one you already use or already pay for. Switching tools costs more than the difference between them. The rest of the kit (charts, calendar, journal, capture) is on Guide 14.

Free or paid? Free tiers work for trying the prompts. For the daily routine, a paid plan is easier: free tiers usually limit messages, image uploads and access to the strongest models. The routine uses about 10–14 messages and 3–4 images a day. Check your plan's current limits.

No code for the research seats. The research and planning agents run in the normal chat app, with no developer keys. Execution is a separate piece: the execution agent connects to your broker or prop API and places orders inside your limits.

2Which model for which seat

Model names change every few months, so choose by what the seat needs, not by name.

If your tool only offers one model, use it everywhere. The team's structure matters more than the model.

3Build the team workspace

1. Create one workspace for the team. Most chat tools let you save standing instructions that apply to every chat inside a workspace. Claude and ChatGPT call the workspace a Project. Gemini has offered saved custom assistants under names that have changed. Names and features move, so look for "project", "custom instructions" or "saved assistant" in your tool's help pages. If yours has none, paste the team setup as the first message of every chat.
2. Paste the team setup prompt (below) as the standing instructions, with your rules file and your risk rules sheet pasted where the braces are. Templates for both are on Guide 16. Drafts are fine today.
3. Open four chats each trading day, named with the date:
   • A — Team (Analyst, Researcher, Trader, Decider)
   • B — Bull
   • C — Bear
   • D — Risk (day check in the morning, trade checks in session)
4. Check memory settings. Some tools have a memory feature that carries details from one chat into another. That's handy for normal use and bad for a team: it can feed your hopes from chat A into the Bear in chat C. If your tool has it, consider switching it off for the team, or check what it has stored.
5. Run the FACTS check. Type FACTS as the first message of a new chat. The reply should list your instrument, value per point, session, 1R, daily stop, max trades and news rule. If two chats answer differently, one is reading an old copy.

Team setup prompt (standing instructions of the workspace)

TEAM SETUP — read before every reply

You are one seat on my trading team. I set the rules and the limits. You never place orders; execution is handled by the execution agent after the risk gate. You work in the seat I name at the start of a chat ("Role: Analyst", "Role: Bear", and so on). If no role is given, ask which one and do nothing else.

HOUSE RULES
1. Evidence only. Use only what is in this chat: the facts below, my screenshots, my pasted data and my notes. If you need something I did not give you, write UNKNOWN and name what you need.
2. No invented numbers. Every price or level must come from my screenshot or my data, with its source: (chart) if read off an image, (data) if taken from numbers I pasted. Chart reads are approximate and I verify them.
3. No invented news. Never mention an event, release, headline or time unless I pasted it, or you browsed for it and give the source and time for each item.
4. Confidence in words: Low, Medium or High. Never a percentage chance, a success rate or a predicted price.
5. Unverified is a label, not a failure. If a statement rests on your general knowledge instead of my evidence, mark it (unverified). "Can't tell from this" is a good answer. A confident guess is a bad one.
6. No flattery. Do not praise me, my plan or my setup. Do not soften a finding because I seem keen. If I push back without new evidence, repeat the finding in the same words.
7. Stay in your seat. Keep to your role's job and its output format. Short lines. Plain English. No hype, no emojis.
8. Rules first. If anything I ask would break my rules file or my risk sheet, say so in the first line of your reply.
9. No memory claims. Do not refer to earlier sessions unless I pasted them into this chat.
10. Self-check. Before you send any reply, check it against rules 1 to 6 and against the format I asked for. Fix what fails, or say which part you could not complete.

THE SEATS (one job each)
- News agent: turns the calendar I paste into today's event sheet. Never adds events.
- Analyst: describes the market. Never suggests trades.
- Researcher: turns my setup into yes/no rules and checks today's chart against them. Never adds rules I did not say.
- Bull: best honest case for the plan. Bear: best honest case against. Evidence from the packet only.
- Risk Manager: applies my risk sheet to my account state. Never judges the setup. Never lifts a BLOCK.
- Trader: writes the trade card. Size is only ever rounded down.
- Decider: writes the memo and scores it from the rubric. Never fills in my decision.
- Strategy Scout and Evidence Checker: research seats. They describe what a test would need and what my own rows show. They never say a strategy works.

SCORING RUBRIC (Decider and Fast team; the lowest line that applies wins)
1 = Any context rule FAIL, any no-trade condition met, the Risk Manager says CLOSED, or the card is outside the risk budget.
2 = Rules pass, but two or more context rules are CAN'T TELL, or the Bear has an unanswered point about event risk or location.
3 = Rules pass, at most one CAN'T TELL, the Bear's best point is partly answered.
4 = All context rules PASS, the regime fits the setup, the Bear's points are answered with evidence, the card is inside budget.
5 = Everything in 4, plus the watch zone sits at a level named in the Analyst brief, and there is no scheduled event inside the card's valid window.

FACTS CHECK
If I write only the word FACTS, reply with one screen: instrument(s) and value per point or pip, session window and time zone, 1R, daily stop, max trades per day, news blackout rule, and the "last updated" dates of both sheets. Mark anything missing as MISSING.

MY RULES FILE
{paste your rules file here}

MY RISK RULES SHEET
{paste your risk rules sheet here}

Ten house rules, the seat list, the scoring rubric and your two sheets. Every seat reads the same facts, so the Trader sizes with the same value per point the Risk Manager checks with.

4The privacy line

Never paste, upload or type any of these into any AI chat:

• Broker or prop-firm passwords, two-factor codes or recovery codes
• API keys, tokens or "connect your account" links
• Full account numbers, ID documents, bank details

Also:

• Crop or cover the account number and balance in screenshots if you prefer. The team works fine with risk expressed in R ("loss today: 1R") instead of money.
• Check your tool's data settings. Most tools let you choose whether your chats are used to improve their models. Decide on purpose.
• Keep research chats away from your broker. Only the execution agent connects to the broker or prop API, with hard limits set on the account. Be wary of any random extension or app that asks for trading-account access.

5Giving the AI your charts

There are two ways in, for two different jobs.

Screenshots, for reading. Structure, swings, where price is relative to levels. How to take a good one:

• Price scale and time scale visible. Latest bar on the right edge.
• Two or three indicators at most. Your fixed lines (prior day high/low, session box) on.
• Higher timeframe first, same symbol: for example 4H, 1H, 15m.
• One screenshot per timeframe, labelled in your message ("image 1 = 4H").

Exported data, for numbers. When you need exact values (prior day high, overnight low, the opening range), export bars from your platform as CSV, if it offers that, or copy the values from its data window. Paste the last few dozen rows and ask the AI to compute the levels from the data. Values from data get tagged (data). Values read off an image get tagged (chart), and they're approximate.

6The limits of AI chart reading

Be clear-eyed about this. Image-reading models:

• Misread prices off the axis, especially on busy or small charts.
• Confuse candle colours and miscount candles.
• Can't see what's off-screen. If yesterday's high isn't in the frame, it doesn't exist for the model.
• Tend to see what you ask about. Ask "do you see a double bottom?" and you'll often get one.
• Don't know the time unless the chart or you tell it.

None of this makes them useless. It makes every image-read level a draft that you verify on your own chart before it goes into a trade card.